Use AI as a second reader of interviews

Goal & fit

Develop a researcher-owned analytic memo and an evidence-linked set of candidate themes. AI suggestions are provisional readings, not findings or substitutes for qualitative judgment.

Suitable when

Exploratory assistance with an approved corpus when a qualitative researcher can read the underlying interviews and challenge summaries. A small pilot can reveal whether suggestions add anything to a human reading.

Not suitable when

Interviews whose consent or sensitivity rules prohibit the proposed processing, or an analysis where no researcher can verify the language and context. Do not use generated themes to replace participant testimony or infer diagnoses, identities or motives.

Inputs to gather

• Consent-compatible transcripts with stable interview and segment IDs, a protected original and a checked working copy.
• A research question, methodological position and reflexive notes about the researcher's relationship to the data.
• A documented data-processing decision and a plan for reviewing suggestions against full interview context.

Suggested workflow

The sequence below is editorial guidance. Numbered references identify sources for tool capabilities, not evidence that this entire workflow was tested.

1. Resolve consent and prepare the working copy

   Check consent, ethics approval and institutional processing requirements before any upload. Remove direct identifiers and review combinations of indirect identifiers. Keep the identity key separately with restricted access. De-identification reduces risk but does not make a rich life narrative anonymous.

2. Read before asking for a model reading

   Read whole interviews and write initial observations, tensions and positionality notes. Decide whether you are using a codebook-oriented approach or an interpretive, reflexive approach; do not silently turn one into the other because a tool offers automatic codes.

3. Pilot a narrow assistive task

   Choose a varied, approved subset and ask for candidate codes or a summary anchored to segment IDs. MAXQDA documents code suggestions and summaries [1]; ATLAS.ti documents document chat and AI coding [2]. Save the exact input selection and available model or feature version. Treat every suggested quote as unverified until matched to the transcript.

4. Compare suggestions with the human reading

   For each suggestion, record the supporting excerpt, surrounding context and your accept, revise or reject decision. Look for flattening of irony, silence, ambiguity and minority experiences. A fluent summary is not evidence that the participant intended the model's interpretation.

5. Build and challenge the interpretation

   Develop themes across interviews rather than promoting frequently repeated labels automatically. Seek disconfirming cases and alternative explanations. Keep an analytic memo showing why a theme is useful, its boundaries and which interviews complicate it. Agreement scores are not a universal validity test for reflexive thematic analysis.

6. Write an auditable account

   Separate participant quotations, AI paraphrases and researcher interpretation. Explain which tasks received AI assistance and which decisions remained human. Preserve a restricted audit trail and share only consent-compatible excerpts; remove unsupported quotations rather than smoothing them into plausible prose.

Alternative tools & routes

Choose by project needs and governance, not by an assumed quality ranking. Current costs, licences and service terms need checking before adoption.

Expected output

An analytic memo, candidate-theme table with segment references and counterexamples, a decision log for AI suggestions, and a disclosure of the assisted tasks. A human-only reading and memo workflow is the comparison route, not a lesser fallback.

Validation checklist

Use these as acceptance checks in your project record. This page does not save completion state.

• Consent and processing approval cover the actual tool and data flow, not simply 'AI' in the abstract.
• Every quotation matches the original transcript verbatim and retains enough context to be meaningful.
• Candidate themes include counterexamples and researcher rationale, not just lists of frequent terms.
• AI paraphrases cannot be confused with participant quotations.
• A reader can trace a reported finding through the memo to the interview evidence without exposing unnecessary personal data.

Privacy & research integrity

• Current retention, training-use, subprocessors and regional hosting terms were not audited. Stop before upload if these are unresolved for your data.
• AI may reproduce stereotypes, mistranslate dialect or suppress disagreement. Model consistency does not establish interpretive validity.
• Avoid replacing lived experience with inferred sentiment. Do not upload credentials, the identity key or unrelated participant records. Retain only the audit material your ethics and retention policy permit.

Linked primary sources

Based on stored primary-source extracts accessed on 15 October 2026. They were re-read for these guides, not freshly fetched or independently reproduced. Features and terms may change.

1. MAXQDA — What is MAXQDA AI Assist?

   Vendor description of summaries, paraphrases and coding support; does not validate this suggested method.

   Source accessed: 2026-10-15

2. ATLAS.ti — official product documentation

   Vendor overview supporting document chat and AI coding capabilities, not accuracy or research outcomes.

   Source accessed: 2026-10-15